1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-gemm/neonfp16arith-ld64.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 
11 #include <assert.h>
12 
13 #include <arm_neon.h>
14 
15 #include <xnnpack/common.h>
16 
17 #include <xnnpack/gemm.h>
18 
19 
xnn_f16_gemminc_minmax_ukernel_6x16__neonfp16arith_ld64(size_t mr,size_t nc,size_t kc,const void * restrict a,size_t a_stride,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,const void * restrict acc,const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f16_gemminc_minmax_ukernel_6x16__neonfp16arith_ld64(
21     size_t mr,
22     size_t nc,
23     size_t kc,
24     const void* restrict a,
25     size_t a_stride,
26     const void* restrict w,
27     void* restrict c,
28     size_t cm_stride,
29     size_t cn_stride,
30     const void*restrict acc,
31     const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
32 {
33   assert(mr != 0);
34   assert(mr <= 6);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(kc % sizeof(__fp16) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41   assert(acc != NULL);
42 
43   const __fp16* a0 = (const __fp16*) a;
44   __fp16* c0 = (__fp16*) c;
45   const __fp16* a1 = (const __fp16*) ((uintptr_t) a0 + a_stride);
46   __fp16* c1 = (__fp16*) ((uintptr_t) c0 + cm_stride);
47   if XNN_UNPREDICTABLE(mr < 2) {
48     a1 = a0;
49     c1 = c0;
50   }
51   const __fp16* a2 = (const __fp16*) ((uintptr_t) a1 + a_stride);
52   __fp16* c2 = (__fp16*) ((uintptr_t) c1 + cm_stride);
53   if XNN_UNPREDICTABLE(mr <= 2) {
54     a2 = a1;
55     c2 = c1;
56   }
57   const __fp16* a3 = (const __fp16*) ((uintptr_t) a2 + a_stride);
58   __fp16* c3 = (__fp16*) ((uintptr_t) c2 + cm_stride);
59   if XNN_UNPREDICTABLE(mr < 4) {
60     a3 = a2;
61     c3 = c2;
62   }
63   const __fp16* a4 = (const __fp16*) ((uintptr_t) a3 + a_stride);
64   __fp16* c4 = (__fp16*) ((uintptr_t) c3 + cm_stride);
65   if XNN_UNPREDICTABLE(mr <= 4) {
66     a4 = a3;
67     c4 = c3;
68   }
69   const __fp16* a5 = (const __fp16*) ((uintptr_t) a4 + a_stride);
70   __fp16* c5 = (__fp16*) ((uintptr_t) c4 + cm_stride);
71   if XNN_UNPREDICTABLE(mr != 6) {
72     a5 = a4;
73     c5 = c4;
74   }
75 
76   do {
77     float16x8_t vacc0x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
78     float16x8_t vacc0x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
79     float16x8_t vacc1x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
80     float16x8_t vacc1x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
81     float16x8_t vacc2x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
82     float16x8_t vacc2x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
83     float16x8_t vacc3x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
84     float16x8_t vacc3x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
85     float16x8_t vacc4x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
86     float16x8_t vacc4x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
87     float16x8_t vacc5x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
88     float16x8_t vacc5x89ABCDEF = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
89 
90     size_t k = kc;
91     while (k >= 4 * sizeof(__fp16)) {
92       const float16x4_t va0 = vld1_f16(a0); a0 += 4;
93       const float16x4_t va1 = vld1_f16(a1); a1 += 4;
94       const float16x4_t va2 = vld1_f16(a2); a2 += 4;
95       const float16x4_t va3 = vld1_f16(a3); a3 += 4;
96       const float16x4_t va4 = vld1_f16(a4); a4 += 4;
97       const float16x4_t va5 = vld1_f16(a5); a5 += 4;
98 
99       const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
100       const float16x8_t vb89ABCDEFc0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
101 
102       #if XNN_ARCH_ARM64
103         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
104         vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c0, va1, 0);
105         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);
106         vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c0, va3, 0);
107         vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c0, va4, 0);
108         vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c0, va5, 0);
109         vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);
110         vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc0, va1, 0);
111         vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc0, va2, 0);
112         vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc0, va3, 0);
113         vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc0, va4, 0);
114         vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc0, va5, 0);
115       #else
116         const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
117         const float16x8_t va1c0 = vdupq_lane_f16(va1, 0);
118         const float16x8_t va2c0 = vdupq_lane_f16(va2, 0);
119         const float16x8_t va3c0 = vdupq_lane_f16(va3, 0);
120         const float16x8_t va4c0 = vdupq_lane_f16(va4, 0);
121         const float16x8_t va5c0 = vdupq_lane_f16(va5, 0);
122 
123         vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
124         vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c0, vb01234567c0);
125         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c0, vb01234567c0);
126         vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c0, vb01234567c0);
127         vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c0, vb01234567c0);
128         vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c0, vb01234567c0);
129         vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c0, vb89ABCDEFc0);
130         vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c0, vb89ABCDEFc0);
131         vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c0, vb89ABCDEFc0);
132         vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c0, vb89ABCDEFc0);
133         vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c0, vb89ABCDEFc0);
134         vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c0, vb89ABCDEFc0);
135       #endif
136       const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
137       const float16x8_t vb89ABCDEFc1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
138 
139       #if XNN_ARCH_ARM64
140         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
141         vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c1, va1, 1);
142         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);
143         vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c1, va3, 1);
144         vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c1, va4, 1);
145         vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c1, va5, 1);
146         vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);
147         vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc1, va1, 1);
148         vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc1, va2, 1);
149         vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc1, va3, 1);
150         vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc1, va4, 1);
151         vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc1, va5, 1);
152       #else
153         const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
154         const float16x8_t va1c1 = vdupq_lane_f16(va1, 1);
155         const float16x8_t va2c1 = vdupq_lane_f16(va2, 1);
156         const float16x8_t va3c1 = vdupq_lane_f16(va3, 1);
157         const float16x8_t va4c1 = vdupq_lane_f16(va4, 1);
158         const float16x8_t va5c1 = vdupq_lane_f16(va5, 1);
159 
160         vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
161         vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c1, vb01234567c1);
162         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c1, vb01234567c1);
163         vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c1, vb01234567c1);
164         vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c1, vb01234567c1);
165         vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c1, vb01234567c1);
166         vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c1, vb89ABCDEFc1);
167         vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c1, vb89ABCDEFc1);
168         vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c1, vb89ABCDEFc1);
169         vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c1, vb89ABCDEFc1);
170         vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c1, vb89ABCDEFc1);
171         vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c1, vb89ABCDEFc1);
172       #endif
173       const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
174       const float16x8_t vb89ABCDEFc2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
175 
176       #if XNN_ARCH_ARM64
177         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
178         vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c2, va1, 2);
179         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);
180         vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c2, va3, 2);
181         vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c2, va4, 2);
182         vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c2, va5, 2);
183         vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);
184         vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc2, va1, 2);
185         vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc2, va2, 2);
186         vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc2, va3, 2);
187         vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc2, va4, 2);
188         vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc2, va5, 2);
189       #else
190         const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
191         const float16x8_t va1c2 = vdupq_lane_f16(va1, 2);
192         const float16x8_t va2c2 = vdupq_lane_f16(va2, 2);
193         const float16x8_t va3c2 = vdupq_lane_f16(va3, 2);
194         const float16x8_t va4c2 = vdupq_lane_f16(va4, 2);
195         const float16x8_t va5c2 = vdupq_lane_f16(va5, 2);
196 
197         vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
198         vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c2, vb01234567c2);
199         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c2, vb01234567c2);
200         vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c2, vb01234567c2);
201         vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c2, vb01234567c2);
202         vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c2, vb01234567c2);
203         vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c2, vb89ABCDEFc2);
204         vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c2, vb89ABCDEFc2);
205         vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c2, vb89ABCDEFc2);
206         vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c2, vb89ABCDEFc2);
207         vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c2, vb89ABCDEFc2);
208         vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c2, vb89ABCDEFc2);
209       #endif
210       const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
211       const float16x8_t vb89ABCDEFc3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
212 
213       #if XNN_ARCH_ARM64
214         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
215         vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c3, va1, 3);
216         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);
217         vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c3, va3, 3);
218         vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c3, va4, 3);
219         vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c3, va5, 3);
220         vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);
221         vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc3, va1, 3);
222         vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc3, va2, 3);
223         vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc3, va3, 3);
224         vacc4x89ABCDEF = vfmaq_lane_f16(vacc4x89ABCDEF, vb89ABCDEFc3, va4, 3);
225         vacc5x89ABCDEF = vfmaq_lane_f16(vacc5x89ABCDEF, vb89ABCDEFc3, va5, 3);
226       #else
227         const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
228         const float16x8_t va1c3 = vdupq_lane_f16(va1, 3);
229         const float16x8_t va2c3 = vdupq_lane_f16(va2, 3);
230         const float16x8_t va3c3 = vdupq_lane_f16(va3, 3);
231         const float16x8_t va4c3 = vdupq_lane_f16(va4, 3);
232         const float16x8_t va5c3 = vdupq_lane_f16(va5, 3);
233 
234         vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
235         vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c3, vb01234567c3);
236         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c3, vb01234567c3);
237         vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c3, vb01234567c3);
238         vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c3, vb01234567c3);
239         vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3);
240         vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0c3, vb89ABCDEFc3);
241         vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1c3, vb89ABCDEFc3);
242         vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2c3, vb89ABCDEFc3);
243         vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c3, vb89ABCDEFc3);
244         vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c3, vb89ABCDEFc3);
245         vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c3, vb89ABCDEFc3);
246       #endif
247 
248       k -= 4 * sizeof(__fp16);
249     }
250     if XNN_UNLIKELY(k != 0) {
251       do {
252         const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
253         const float16x8_t va1 = vld1q_dup_f16(a1); a1 += 1;
254         const float16x8_t va2 = vld1q_dup_f16(a2); a2 += 1;
255         const float16x8_t va3 = vld1q_dup_f16(a3); a3 += 1;
256         const float16x8_t va4 = vld1q_dup_f16(a4); a4 += 1;
257         const float16x8_t va5 = vld1q_dup_f16(a5); a5 += 1;
258 
259         const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
260         const float16x8_t vb89ABCDEF = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
261 
262         vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
263         vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1, vb01234567);
264         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);
265         vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3, vb01234567);
266         vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4, vb01234567);
267         vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5, vb01234567);
268         vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0, vb89ABCDEF);
269         vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1, vb89ABCDEF);
270         vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2, vb89ABCDEF);
271         vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3, vb89ABCDEF);
272         vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4, vb89ABCDEF);
273         vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5, vb89ABCDEF);
274 
275         k -= sizeof(__fp16);
276       } while (k != 0);
277     }
278 
279     const float16x8_t vscale = vld1q_dup_f16((const __fp16*) &params->scale);
280     vacc0x01234567 = vmulq_f16(vacc0x01234567, vscale);
281     vacc1x01234567 = vmulq_f16(vacc1x01234567, vscale);
282     vacc2x01234567 = vmulq_f16(vacc2x01234567, vscale);
283     vacc3x01234567 = vmulq_f16(vacc3x01234567, vscale);
284     vacc4x01234567 = vmulq_f16(vacc4x01234567, vscale);
285     vacc5x01234567 = vmulq_f16(vacc5x01234567, vscale);
286     vacc0x89ABCDEF = vmulq_f16(vacc0x89ABCDEF, vscale);
287     vacc1x89ABCDEF = vmulq_f16(vacc1x89ABCDEF, vscale);
288     vacc2x89ABCDEF = vmulq_f16(vacc2x89ABCDEF, vscale);
289     vacc3x89ABCDEF = vmulq_f16(vacc3x89ABCDEF, vscale);
290     vacc4x89ABCDEF = vmulq_f16(vacc4x89ABCDEF, vscale);
291     vacc5x89ABCDEF = vmulq_f16(vacc5x89ABCDEF, vscale);
292 
293     const float16x8_t vmax = vld1q_dup_f16((const __fp16*) &params->max);
294     vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
295     vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
296     vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);
297     vacc3x01234567 = vminq_f16(vacc3x01234567, vmax);
298     vacc4x01234567 = vminq_f16(vacc4x01234567, vmax);
299     vacc5x01234567 = vminq_f16(vacc5x01234567, vmax);
300     vacc0x89ABCDEF = vminq_f16(vacc0x89ABCDEF, vmax);
301     vacc1x89ABCDEF = vminq_f16(vacc1x89ABCDEF, vmax);
302     vacc2x89ABCDEF = vminq_f16(vacc2x89ABCDEF, vmax);
303     vacc3x89ABCDEF = vminq_f16(vacc3x89ABCDEF, vmax);
304     vacc4x89ABCDEF = vminq_f16(vacc4x89ABCDEF, vmax);
305     vacc5x89ABCDEF = vminq_f16(vacc5x89ABCDEF, vmax);
306 
307     const float16x8_t vmin = vld1q_dup_f16((const __fp16*) &params->min);
308     vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
309     vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
310     vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);
311     vacc3x01234567 = vmaxq_f16(vacc3x01234567, vmin);
312     vacc4x01234567 = vmaxq_f16(vacc4x01234567, vmin);
313     vacc5x01234567 = vmaxq_f16(vacc5x01234567, vmin);
314     vacc0x89ABCDEF = vmaxq_f16(vacc0x89ABCDEF, vmin);
315     vacc1x89ABCDEF = vmaxq_f16(vacc1x89ABCDEF, vmin);
316     vacc2x89ABCDEF = vmaxq_f16(vacc2x89ABCDEF, vmin);
317     vacc3x89ABCDEF = vmaxq_f16(vacc3x89ABCDEF, vmin);
318     vacc4x89ABCDEF = vmaxq_f16(vacc4x89ABCDEF, vmin);
319     vacc5x89ABCDEF = vmaxq_f16(vacc5x89ABCDEF, vmin);
320 
321     if XNN_LIKELY(nc >= 16) {
322       vst1q_f16(c0, vacc0x01234567);
323       vst1q_f16(c0 + 8, vacc0x89ABCDEF);
324       c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
325       vst1q_f16(c1, vacc1x01234567);
326       vst1q_f16(c1 + 8, vacc1x89ABCDEF);
327       c1 = (__fp16*) ((uintptr_t) c1 + cn_stride);
328       vst1q_f16(c2, vacc2x01234567);
329       vst1q_f16(c2 + 8, vacc2x89ABCDEF);
330       c2 = (__fp16*) ((uintptr_t) c2 + cn_stride);
331       vst1q_f16(c3, vacc3x01234567);
332       vst1q_f16(c3 + 8, vacc3x89ABCDEF);
333       c3 = (__fp16*) ((uintptr_t) c3 + cn_stride);
334       vst1q_f16(c4, vacc4x01234567);
335       vst1q_f16(c4 + 8, vacc4x89ABCDEF);
336       c4 = (__fp16*) ((uintptr_t) c4 + cn_stride);
337       vst1q_f16(c5, vacc5x01234567);
338       vst1q_f16(c5 + 8, vacc5x89ABCDEF);
339       c5 = (__fp16*) ((uintptr_t) c5 + cn_stride);
340 
341       a0 = (const __fp16*) ((uintptr_t) a0 - kc);
342       a1 = (const __fp16*) ((uintptr_t) a1 - kc);
343       a2 = (const __fp16*) ((uintptr_t) a2 - kc);
344       a3 = (const __fp16*) ((uintptr_t) a3 - kc);
345       a4 = (const __fp16*) ((uintptr_t) a4 - kc);
346       a5 = (const __fp16*) ((uintptr_t) a5 - kc);
347 
348       nc -= 16;
349     } else {
350       if (nc & 8) {
351         vst1q_f16(c0, vacc0x01234567); c0 += 8;
352         vst1q_f16(c1, vacc1x01234567); c1 += 8;
353         vst1q_f16(c2, vacc2x01234567); c2 += 8;
354         vst1q_f16(c3, vacc3x01234567); c3 += 8;
355         vst1q_f16(c4, vacc4x01234567); c4 += 8;
356         vst1q_f16(c5, vacc5x01234567); c5 += 8;
357 
358         vacc0x01234567 = vacc0x89ABCDEF;
359         vacc1x01234567 = vacc1x89ABCDEF;
360         vacc2x01234567 = vacc2x89ABCDEF;
361         vacc3x01234567 = vacc3x89ABCDEF;
362         vacc4x01234567 = vacc4x89ABCDEF;
363         vacc5x01234567 = vacc5x89ABCDEF;
364       }
365       float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
366       float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
367       float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);
368       float16x4_t vacc3x0123 = vget_low_f16(vacc3x01234567);
369       float16x4_t vacc4x0123 = vget_low_f16(vacc4x01234567);
370       float16x4_t vacc5x0123 = vget_low_f16(vacc5x01234567);
371       if (nc & 4) {
372         vst1_f16(c0, vacc0x0123); c0 += 4;
373         vst1_f16(c1, vacc1x0123); c1 += 4;
374         vst1_f16(c2, vacc2x0123); c2 += 4;
375         vst1_f16(c3, vacc3x0123); c3 += 4;
376         vst1_f16(c4, vacc4x0123); c4 += 4;
377         vst1_f16(c5, vacc5x0123); c5 += 4;
378 
379         vacc0x0123 = vget_high_f16(vacc0x01234567);
380         vacc1x0123 = vget_high_f16(vacc1x01234567);
381         vacc2x0123 = vget_high_f16(vacc2x01234567);
382         vacc3x0123 = vget_high_f16(vacc3x01234567);
383         vacc4x0123 = vget_high_f16(vacc4x01234567);
384         vacc5x0123 = vget_high_f16(vacc5x01234567);
385       }
386       if (nc & 2) {
387         vst1_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
388         vst1_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpret_u32_f16(vacc1x0123), 0); c1 += 2;
389         vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_f16(vacc2x0123), 0); c2 += 2;
390         vst1_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpret_u32_f16(vacc3x0123), 0); c3 += 2;
391         vst1_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpret_u32_f16(vacc4x0123), 0); c4 += 2;
392         vst1_lane_u32(__builtin_assume_aligned(c5, 1), vreinterpret_u32_f16(vacc5x0123), 0); c5 += 2;
393 
394         vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
395         vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
396         vacc2x0123 = vext_f16(vacc2x0123, vacc2x0123, 2);
397         vacc3x0123 = vext_f16(vacc3x0123, vacc3x0123, 2);
398         vacc4x0123 = vext_f16(vacc4x0123, vacc4x0123, 2);
399         vacc5x0123 = vext_f16(vacc5x0123, vacc5x0123, 2);
400       }
401       if (nc & 1) {
402         vst1_lane_f16(c0, vacc0x0123, 0);
403         vst1_lane_f16(c1, vacc1x0123, 0);
404         vst1_lane_f16(c2, vacc2x0123, 0);
405         vst1_lane_f16(c3, vacc3x0123, 0);
406         vst1_lane_f16(c4, vacc4x0123, 0);
407         vst1_lane_f16(c5, vacc5x0123, 0);
408       }
409 
410       nc = 0;
411     }
412   } while (nc != 0);
413 }
414